Skip to contents

For every ply in a batch of self-play games this records the heuristic evaluation from the perspective of the position's player-to-move, plus the eventual outcome for that player (+1 win, -1 loss, 0 draw). The resulting tibble feeds mxo_fit_calibration().

Usage

mxo_make_calibration_data(
  n_games = 100L,
  policy_x = mxo_policy("random"),
  policy_o = mxo_policy("random"),
  config = mxo_config_default(),
  seed = NULL
)

Arguments

n_games

Integer scalar, number of self-play games to run.

policy_x, policy_o

mxo_policy objects. Default: two random policies.

config

A config list (see mxo_config_default()).

seed

Optional integer base seed.

Value

A tibble with columns score (dbl, mover-perspective evaluation) and outcome (int, in {-1, 0, 1} for mover-loss / draw / mover-win).

Examples

cd <- mxo_make_calibration_data(
  n_games = 3L, config = mxo_config_default(n = 3L, ply_cap = 12L),
  seed = 1L
)
table(cd$outcome)
#> 
#> -1  0  1 
#>  5 24  5